Faculty Of Computer Science Graduation Project 2020 - 2022

Permanent URI for this collectionhttp://185.252.233.37:4000/handle/123456789/5002

Browse

Now showing 1 - 19 of 19

Applying Deep Learning to Track Food Consumption and Human Activity for Non-intrusive Blood Glucose Monitoring
(October University For Modern Sciences and Arts, 2022) Mohamed, Mohamed Amr Samir
Diabetes management is a very important eld of study that focuses on improving the lives of peoples living with diabetes. Diabetes Mellitus (DM), a common human condition characterised by hyperglycemia, includes a number of severe complications. In addition, hypoglycemia which is the decrease in blood glucose levels is linked to catastrophic brain failure and death. In this project, we look at a range of relevant research to get a better understanding of some of the systems and concepts that can help in constructing an autonomous glucose monitoring system, including deep learning approaches. Then we proceed to introduce our system that combines non-intrusive Continuous Glucose Monitoring (CGM) devices with blood glucose forecasting, classi cation of meal images, and human activity recognition. We apply deep learning to predict the patient's future glucose levels by combining several models that we have tested. Predictions are made by learning the patterns of blood glucose changes throughout the day of the diabetic patient with the help of meal detection and human activity recognition, allowing the system to estimate future blood glucose levels and warn against hyperglycemia and hypoglycemia even when CGM device becomes unwearable. We have conducted four experiments, one of which is a comparison of deep learning models for food classi cation and human activity recognition, and a couple more experimental results obtained by constructing a glucose forecasting system that combines our deep learning models.
Automatic Short Answer Grading
(October University For Modern Sciences and Arts, 2022) Mohamed Saeed, Mostafa
Much research has been done on automatic grading of student answers since 1966, and it was divided into short answer grading and essay scoring. Our main target is on the short answer-grading task. Therefore, we have implemented two modules an Ensemble-model that is based on similarity algorithms and Neural Network module using sentence embedding pre-trained models. In the first module, we are implementing some text similarity algorithms on Texas dataset. These text similarity algorithms are classified into string similarity using Abydos package that contains 168 string similarity algorithms, semantic similarity (corpus and knowledge-based) and different deep learning embedding models similarity (transformers). Different experiments were done by testing them separately and combining them too, to propose our new model and methodology to achieve the maximum correlation that can be produced from this task using this module which was 65.14%. The neural network module, which used the T5 sentence embedding pre-trained model, reached 92.80 % correlation score, which is significantly better than the other module and had the greatest correlation result when compared to other studies.
Automatic Summarization of Scientific Articles
(October University For Modern Sciences and Arts, 2022) Waly, Rana Reda
The process of scientific research starts with studying the state of the art, and this means an infinite number of publications. Hence, Automatic Summarization of Scientific Articles will help scholars, researchers, or anyone interested in a specific topic by summarizing and shortening the articles they have to read, this will save them a lot of time allowing them to read more articles and gather more information. In conclusion, the proposed solution is divided into two approaches the Extractive approach and the Abstractive approach. For the Extractive approach, different embedding techniques and different approaches were used. But, the one that gave the best results was using GloVe as an embedding technique, and as for the number of sentences to be extracted after many experiments a new approach was applied and it gave a promising result with 0.1244 Rouge-2 f-measures. And as for the Abstractive approach, different pre-trained models were used to solve this problem, but the one that gave the highest results was the GPT-2 XL pre-trained model with 0.38733 Rouge-2 f-measure. As for the dataset used was the CL-SciSumm 2019.
A Blockchain-based System for Electronic Health Records
(October University For Modern Sciences and Arts, 2022) Ahmed Abdo, Seif
Secure storage and Privacy protection of medical data have always been a crucial is- sue for the population during interacting with medical services but now the presence of blockchain technology brings a new idea to solve this problem as it is decentralized and immutable. The main idea of this project is to record the medical data like diagnosis, Medicine prescribed, blood group and more comprehensive data.
Creating Digital Painting Lighting Effects via RGB-space Geometry and Deep Learning
(October University For Modern Sciences and Arts, 2022) Elkhateeb, Marawan Aboelseoud
A way for generating illumination in a digital painting is represented. Two methods were approached for relighting pictures, first was a deep learning approach and the second was a traditional approach that is based on a critical observation. Painters paint lighting effects with multiple overlapping strokes; therefore, pixels with rich stroke historical past tend to accumulate more lighting strokes. Based on this identification, a technique was created that takes color geometry to assess the density of strokes in a painting, and then generates innovative lighting effects by replicating painters’ coarse- to-fine workflow. A wave transform is used to produce coarse illumination, which are subsequently retouched into usable lighting effects based on the stroke density of the source artwork. In addition, we urge artists to create lighting effects for comparison. Lighting artist, who is encouraged to paint lighting effects similar to the used approach’s results, and another non-lighting specialized artist, who is requested to do their best to relight the original image according to artistic vision. All results are then compared to evaluate the accuracy of the used approaches.
Dental implant recognition and classi cation with Convolutional Neural Network
(October University For Modern Sciences and Arts, 2022) Sadek, Andrew Ayman Edward
The dental implants market was worth over USD 7,222 million in 2020, and it's pre- dicted to grow to USD 11,801 million by 2026, with a compound annual growth rate of 8.6 % over the forecast period of 2021-2026. This demonstrates that the number of dental implants will dramatically increase by 2026. These contributions will create a problem for dentists all over the globe in identifying the type of implant and getting the manufacturer's company contacts. This thesis will discuss how the presented system identi ed four types of implants with acceptable accuracies. Three CNN models used are: VGG16, Xception, and ResNet50V2 were applied with transfer learning to train the models on the implants.
Designing Potential Inhibitors For SARS-CoV-2 Main Protease Using Deep Learning
(October University For Modern Sciences and Arts, 2022) Hassan, Adham Khaled
in this work we are trying to speed up the process of finding a cure for SARS-CoV-2 since SARS-CoV-2 have impacted our society due to the global pandemic which have affected the education, economy, world heath care and deaths caused by the virus due to the long time taken by drug discovery pipeline which is between 10 to 12 yeas for a drug to be develop and the enormous cost of the drug discovery pipeline and the low success rate of drug passing the FDA approve in the have motivated us to design a deep learning solution for designing a drug in fraction of the time required and for the current event done by the SARS-CoV-2 the target for this proposed solution will be SARS-CoV-2 the proposed solution is consist of two model one for generative molecules and the other for predicting the affinity of the molecule toward SARS-CoV-2 the proposed solution achieved a generation of molecules with average affinity of -9.8 and a prediction of accuracy of 98.1625% toward SARS-CoV-2, the proposed solution could reduce the drug discovery pipeline which is between 10 to 12 yeas to only 1 to 3 yeas for any novel virus.
Estimation of Glucose Levels Using Smartwatches containing ECG Sensors
(October University For Modern Sciences and Arts, 2022) Maged Sabry, Youssef
The goal of this project is to evaluate various machine learning and deep learning models on patients from the D1NAMO dataset to predict blood glucose level based on heart rate values and additional features extracted from the smart watch. These features will then be processed with feature extraction techniques to identify the features that are most strongly correlated with the glucose values. The suggested system will be patient-dependent and operate in the following manner: each patient will need to get the two devices which are the CGM and the smart watch, while the CGM device measures glucose levels over the course of two weeks, the smartwatch simultaneously measures heart-rate measurements, and both devices will be calibrated together simultaneously. After The data is done being extracted from the two devices mentioned, the data will be taken to be processed, they will be processed by removing any noise that might interfere with the ECG signal, also due to the time measuring differences between the two devices, moving average techniques needed to be applied on the ECG readings in order to be compatible with the glucose’s time readings. This technique is applied because the CGM device measures the glucose values every 5 minutes while the smartwatch measures heart rate every millisecond. The Processed and calibrated data will now be used as a training dataset for the different models that will be used throughout this project. It was found that among the different machine learning models that were used, the extra trees regressor model scored the least MSE (mean squared error) and among the deep learning, neural network models, the ANN (artificial neural network) model scored the least MSE.
Football Match Events expressing for visually and hearing Impaired
(October University For Modern Sciences and Arts, 2022) Badr Abdelqader, Mohammad
Football is the most viewed sport in the world. Many visually and hearing impaired individuals deserve to keep up with matches like normal people, so a players classification model has been implemented to classify players and referees. As well as a ball detection model that would help to detect the ball positions on the field in addition to an action recognition algorithm that would help to get the current state of the ball in order to perform a simulation for the players and ball movement on a 2D pitch using perspective transformation. 9
Forensic Document Examination For Handwritten Signatures Using Deep Learning
(October University For Modern Sciences and Arts, 2022) Tarek Ibrahim, Omar
Forgery is a type of fraud de ned as the act of forging a copy or an imitation of a document, signature, or banknote which is considered a form of illegal criminal activity. In this project, we are focusing on developing an application that aims to help forensic examiners in the process of identi cation and detection of handwritten signature forgeries inside documents in addition to the identi cation of possible types of signature forgery methods. The proposed system uses contemporary methods that utilize a deep learning approach of CNNs (Convolutional Neural Networks) based on both binary, and categorical image classi cation methodologies to help forensic examiners measure the genuineness of handwritten signatures. For the image classi cations, we have considered using a number of ve di erent CNN models which are: VGG-16, ResNet50, Inception-v3, Xception, and our 2DCNN model. The purpose of using these di erent CNN models is to conduct a comparative study between each of the models to help determine which model is more e cient at identifying images containing text data of similar resemblances, as to the nature of handwritten signatures. Upon comparing these CNN models, we concluded that the ResNet50 model was able to reach the highest score at identifying handwritten signatures with an accuracy of 82.3% and 86% when tested on datasets of 300 images and 140 images respectively. The system aims to provide all functionalities which are required by current standards through the use of robust system architecture, design patterns, and various techniques to build a practical application that follows an expandable structure.
Human Activity Recognition in car workshop
(October University For Modern Sciences and Arts, 2022) Magdy Tawfik, Omar
Human activity recognition has become so widespread in recent times. Due to the modern advancements of technology, it has become an important solution to many problems in various fields such as medicine, industry, and sports. And this subject got the attention of a lot of researchers. Along with problems like wasted time in maintenance centers, we proposed a system that extract worker poses from videos by using pose classification. In this paper, we have tested two algorithms to detect worker activity. This system aims to detect and classify positive and negative worker’s activities in car maintenance centers such as (changing the tire, changing oil, using the phone, standing without work). We have conducted two experiments, the first experiment was for comparison between algorithms to determine the most accurate algorithm in recognizing the activities performed. The experiment was done using two different algorithms (1 dollar recognizer and Fast Dynamic time warping) on 3 participants in a controlled area. The one-dollar recognizer has achieved a 97% accuracy with compared to the fastDTW with 86%. The second experiment was conducted to measure the performance of one-dollar algorithm with different participants. The results show that 1 dollar recognizer achieved an accuracy of 95% when tested on 10 different videos.
Image Inpainting Using Deep Learning
(October University For Modern Sciences and Arts, 2022) Mohamed Marie, Abdelrahman
Image became one of the most used form of data we use in our daily life, so as a result image inpainting became one of the most important topics in image processing. Removing an object is very challenging method as it depends on the complexity of the scene, position and the size of the unwanted object. However, deep learning methods have shown great promise in not only providing excellent results in dealing with the most complex scenes, but also the completed region with high quality pixels which makes the whole scene seems to be realistic. This documentation goes through introducing the idea of image inpainting, and discusses the its importance. It explores the used approaches to achieve it, as well as some previous works for such projects and the evaluation of their methods. Moreover, it covers the full implementation details from the very start to the very end of the of the project including the deep learning algorithm we used and the steps of building it. Showing and comparing some results of different models using different evaluation techniques. It also includes the issues we faced during the project and how we solved these problems. Finally, it shows the intended future work for potential modification and improvements for the project
IRats:Intelligent system for rat behavior analysis
(October University For Modern Sciences and Arts, 2022) Naguib, Andrew Zaky
Rats' behavior analysis is fundamentally crucial in the medical eld and the pharmaceutical industry. Drugs and chemical compounds are given to mice and rats to measure their therapeutic e ects. If the drug's e ect is promising, it will be investigated in humans. Data extraction from experiments and classifying the trajectory of rats and mice will help scienti c researchers speed up the development of new drugs for the community. We used image processing and computer vision to analyze two of the eminent rat behavior analysis experiments, which are the Morris Water Maze (MWM), and Open Field experiments. In the current work, we implemented two experiments, the rst was testing more than one object detection and tracking algorithm on rats and mice behavior analysis. Second, we used three tracking algorithms, which were $P Point-Cloud Recognizer, Dynamic Time Warping (DTW), and FastDTW, to track the rat's path and classify its movements. The result of the rst experiment was that a combination of CSRT (Channel and Spatial Reliability Tracking) and optical ow sparse would be more accurate in extracting the data. In addition, the second experiment showed the highest tracking algorithm was FastDTW, with an accuracy of 81%
Multi-modalities Analysis In Profiled Learning
(October University For Modern Sciences and Arts, 2022) Nady Shoukry, Mario
In the context of the modern development of informational technologies, there is a great impact on education in all aspects such as variety and quality. Moreover, student center approaches become the main goal of many institutions all over the world and this can lead us to Adaptive Learning. Adaptive learning is mainly focusing on enhancing the student whether in the process of learning or even in the assessment. This project aims to take some di erent modalities from the learner and predict his/her result in the exam. We performed an experiment research in which 53 students, ranging in age from 18 to 22, solved an English exam. The study explored the possibility of adding several student modalities such as eye gazing, facial expressions, and mouse movements. The analysis of the collected dataset shows that adding student features can e ectively predict his assessment score. We used di erent regression models to predict the score of the student based on his features.
Prediction of Hypotension in Hemodialysis Sessions
(October University For Modern Sciences and Arts, 2022) Adel Gheith, Salma
There are 4.35 million people who receive the hemodialysis treatment or undergoing a transplant as they had suffered from the session’s duration or the session’s complications. Therefore, this research aims to solve a daily nephrology problem. The main problem is that how a hemodialysis patient suffers for several hours multiple time per week and may have complications during the session which may lead to severe risks. Hypotension is one of the critical problems that faces any dialysis patient where there is a sudden drop in the patient’s blood pressure. According to nephrologists and previous research papers showed how difficult it was to treat and detect the occurrence of hypotension. In addition to, hypotension occurrence has consequences that may lead to death. Therefore, hypotension should be rapidly recognized to avoid any tragic consequences. To achieve this target, a few steps were needed. The research claims to have different approach to handle the hypotension. The first approach is hypotension prediction. Firstly, the data set is required to be from dialysis session to monitor the patient records from the dialysis machine. After preparing the artificial neural network (ANN) model to be able to predict, it needs some adjustment to make it efficient to use and a powerful model to handle the intradialytic event. The second approach is how to handle the occurrence of hypotension through the session and to decide as soon as the blood pressure drops. Therefore, the fuzzy control system is considered as one of the powerful tools to handle similar scenarios as it’s considered as biofeedback system. The second approach depends on different dataset and only few inputs to decide the blood filtration rate of the hemodialysis machine. In conclusion, the 2 approaches aim to handle a major problem that occurs daily in the dialysis units around the world.
Shrimp disease classification and behavior detection
(October University For Modern Sciences and Arts, 2022) Ashraf Wahideldin Hussein, Abdelaziz
Shrimp is a very important aquaculture species in the world. Usually some of viruses affect them. Over the past fifty years, there has been a steady increase in shrimp production worldwide. shrimp production reached 5.5 tons in 2021...
Software Thesis Document for project Sentiment Analysis for Customers Reviews in an E-Commerce Website
(October University For Modern Sciences and Arts, 2022) Megahid, Ahmad Abdallah
The main idea of this project is to analyze customers reviews on an e-commerce website for second-hand apparel. Data analysis is the science discovering out supporting decision-making, useful information, and informing conclusion. It is done by inspecting, discovering, modeling, cleansing, and transforming data. Data analysis here is going to be performed on e-commerce website software system, to know the degree of customer’s satisfaction.
Swimming System Tracker for enhancing Butterfly stroke
(October University For Modern Sciences and Arts, 2022) Tamer, Salma
Swimming is a lifelong beneficial activity. It is an excellent training since it requires you to move your entire body against the water’s resistance; however, by the time, these movements may not be in a right way. In addition, the wrong movements may lead to many pains such as shoulder pain, elbow pain and lower back pain especially in difficult strokes. The coach is the one who instructs the swimmers and tell them which is incorrect, and which is correct. However, he can’t recognize all the incorrect movements, so this needs an instructor who can see all the stroke’s mistakes. Hence our proposed system, which uses machine learning techniques, utilizes four different models which are Long short-term memory (LSTM), k- nearest neighbor (Knn), for time series 1-$ recognizer and Dynamic time wrapping (DTW) to detect the incorrect butterfly stroke. The system uses an accelerometer and gyroscope sensors to detect and evaluate correct and Incorrect swimming patterns in butterfly stoke. In addition to attaching a mobile application to the swimmer’s wrist which gathers all data which allows the coach and the swimmer to know the incorrect strokes such as lifting the head too high, sweeping out after hand entry, and bending the arm. Dynamic time wrappig(DTW) achieved the best accuracy among all classifiers which are an average of 80.5%. The system’s main goal is to use an accelerometer and gyroscope from a mobile sensors to detect and evaluate correct and incorrect swimming patterns in butterfly stoke. We were able to track swimmers’ strokes underwater and therefore aid swimming coaches. When an improper movement is recognised, the system alerts both the swimmer and the coach. The key aim behind the system is to detect and analyze incorrect butterfly swimming style patterns with mobile application attached to their wrist which collects patterns of a swimmer’s stream which enables it to detect the incorrect strokes, count the correct and incorrect strokes, count the total strokes which the swimmer made in specific distance.
Traffic Sign Recognition For Autonomous Driving
(October University For Modern Sciences and Arts, 2022) Mattar, Mahmoud Mohamed
According to world health organization statistics done in 2021, 1.3 million people die yearly as a result of road traffic crashes. Undoubtedly, the fetal injuries are affecting people with age range 5 to 29 which shortly leads to death. recently technology had contributed to reduce the number of crashes by using different applications of computer vision. one of computer vision applications is Traffic sign recognition system (TSRS). clearly, TSRS is a significant partition of intelligent transportation system (ITS). in addition to its ability to identify traffic signs accurately, which effectively can improve the driving safety. This project focuses on traffic sign recognition technique using deep learning, which mainly aims to detect and classify all traffic signs using convolutional neural network (CNN).

Browse

Browsing Faculty Of Computer Science Graduation Project 2020 - 2022 by Title